Adonis0123/adonis-skills / skills/chrome-dev-mcp
chrome-dev-mcp
VerifiedUse this skill when the user invokes /chrome-dev-mcp or asks for Chrome DevTools MCP, CDP, list_pages/select_page, UXC packaging for Chrome DevTools, DOM snapshots, Console, Network, Performance, Lighthouse, heap analysis, browser-internal debugging, connection recovery, or correct-Chrome validation across Claude Code, Codex, Grok Build/Grok002, Cursor CLI, Hermes, or WorkBuddy. Use it when the evidence is Chrome-specific (Lighthouse, a full trace, heap data) or the caller has no page session yet; a page already open in ego-browser keeps its own Console and request diagnosis. Also use it as the verification step right after the user or this session changes a web UI or asks to accept or check a page: collect a screenshot, Console, and Network evidence before calling it done. Do not use it for ordinary navigation, form filling, scraping, or desktop UI unless browser-internal signals or UI-change acceptance are required. A passing readiness or list_pages check proves the connection only, never the change.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.4 KB |
| evals/evals.json | config | 14.7 KB |
| evals/trigger-eval.json | config | 3.3 KB |
| references/host-verification.md | markdown | 5.9 KB |
| references/maintenance.md | markdown | 4.5 KB |
| references/profile-identity.md | markdown | 1.6 KB |
| references/uxc-facade.md | markdown | 7.5 KB |
| scripts/configure-local.zshexecutable | script | 1.8 KB |
| scripts/ensure-connection.zshexecutable | script | 2.6 KB |
| scripts/host-acceptance.zshexecutable | script | 6.9 KB |
| scripts/install-uxc.zshexecutable | script | 4.0 KB |
| scripts/lib/load-config.zshexecutable | script | 438 B |
| scripts/lib/readiness-envelope.zsh | script | 1.5 KB |
| scripts/lib/uxc-link-contract.zsh | script | 1.5 KB |
| scripts/lib/uxc-owned-binary.zsh | script | 2.8 KB |
| scripts/lib/uxc-release.zshexecutable | script | 642 B |
| scripts/setup-uxc-link.zshexecutable | script | 3.1 KB |
| scripts/test/host-acceptance.test.mjs | script | 5.3 KB |
| scripts/test/readiness-envelope.test.mjs | script | 2.7 KB |
| scripts/test/ui-acceptance.test.mjs | script | 5.2 KB |
| scripts/uxc-readiness.zshexecutable | script | 2.7 KB |
references/host-verification.md
View on GitHub# Host verification
Read this reference only for configuration repair, skill discovery, native compatibility, or multi-host acceptance. Begin ordinary browser debugging with shared CLI readiness.
## Separate the layers
Verify the shared skill view and the host's eager-native state separately. On the normal path, `chrome-devtools` must be disabled or absent while `chrome-dev-mcp` remains discoverable.
Examples for hosts that expose configuration inspection:
```bash
claude mcp get chrome-devtools # expect absent
codex mcp list # expect disabled
grok mcp list # expect disabled
cursor-agent mcp list # expect absent (Cursor CLI)
hermes mcp list # expect disabled
```
Then run `zsh scripts/uxc-readiness.zsh` twice. Require `SHARED_TRANSPORT=OK` and require the second call to report `DAEMON_SESSION_REUSED=YES`. Discovery never proves that the shared child attached to the intended Chrome; sanitized readiness does.
For fan-out checks, compare the OS process tree with `uxc daemon sessions`. Count actual Chrome DevTools MCP server processes separately from npm launchers, watchdogs, Chrome renderers and Chrome Helper processes. Run concurrent read-only calls from different project directories and verify the same single shared child remains afterward; `daemon_session_reused=true` alone does not rule out other native children.
If duplicate native children remain, trace each parent to its host and check current user/project/plugin registration. Removing a registration does not stop an already-open session. Within an authorized cleanup, terminate only positively identified legacy MCP processes after confirming they are not the shared child; preserve browser windows and unrelated sessions. Never use a broad name-based kill or a background reaper. Recheck for respawn and fix the owning registration if it returns. Do not promise a permanent global process cap: a separate native registration can still start its own server outside this facade.
## Run optional runtime-lab checks
When `CHROME_DEV_MCP_LAB_DIR` names a trusted checkout, run from that directory:
```bash
./bin/doctor
./bin/test-startup
./bin/test-hosts
```
Keep cold-start tests isolated. Never close an existing browser to simulate recovery.
## Run the real acceptance
Run the bundled script from this skill's directory. It starts one fresh headless session per host in parallel and prints one `HOST=... RESULT=...` line each; page URLs are checked privately and never printed:
```bash
zsh scripts/host-acceptance.zsh # claude, codex, grok, cursor
zsh scripts/host-acceptance.zsh --host codex --keep-raw
```
Each host may run only `zsh scripts/uxc-readiness.zsh` and `chrome-dev-mcp-cli`; no host gets an approve-everything flag. Claude Code and Grok get narrow `--allowedTools`/`--allow` rules, Codex keeps the `workspace-write` sandbox with network and `~/.uxc` added, and Cursor CLI runs in its default mode. Cursor CLI's default headless mode rejects every shell call, so that host reports FAIL without reaching the shared CLI. Pass `--cursor-auto-review` only after the user explicitly approves it: Cursor's server-side classifier then approves calls it judges safe, which cannot be narrowed to the two commands above, so it is wider than the other hosts' rules though narrower than `--force`/`--yolo` ([run modes](https://cursor.com/docs/agent/security/run-modes)). Pass `--bypass` only after the user explicitly approves approve-everything sessions; it uses `--permission-mode bypassPermissions` (Claude Code), `--dangerously-bypass-approvals-and-sandbox` (Codex), `--always-approve` (Grok) and `--force` (Cursor CLI), which removes the narrow rules above for that run. When the calling agent's own permission layer blocks launching other agent CLIs, ask the user to run the script (for example with `! zsh <skill-dir>/scripts/host-acceptance.zsh` in Claude Code) or to allow that exact command; do not retry through another host or wrapper.
The script runs this prompt in every session:
```text
只读验收:必须通过 chrome-dev-mcp skill 的共享 CLI 真实调用 list_pages 一次;不要导航、点击,也不要输出页面 URL、标题或正文。成功只回答 CHROME_DEV_MCP_READY,共享调用失败只回答 CHROME_DEV_MCP_FAIL。不要自动回退到 native MCP。
```
Require all five signals:
1. The shared MCP handshake succeeds.
2. Shared CLI tools are discovered.
3. The model issues a real `list_pages` call.
4. The final sentinel agrees with the tool trace.
5. Captured output contains no page URL, title, body, cookie, or token.
Report provider or quota failures separately from shared transport failures. Test native compatibility only when explicitly requested; label it because it adds a per-session runtime.
## Use explicit native compatibility only as rollback
Treat `NATIVE_COMPAT_REQUIRED` as a blocker, not permission to change configuration. After user approval, restore or enable only the named host's saved `chrome-devtools` entry, use a fresh compatibility session for its real tool call, then return that host to disabled or absent. Do not rewrite unrelated MCP entries, do not use a moving `@latest` registration, and do not assume an already-open session hot-loads the restored tool.
## Report multi-host acceptance
Report each host independently:
| Host | Skill discovery | Shared transport | Real `list_pages` | Native compat | Result |
| ---- | --------------- | ---------------- | ----------------- | ------------- | ------ |
Use `VERIFIED` only after the real shared tool call. Leave native compatibility `NOT_USED` on the healthy default path; otherwise report `UNVERIFIED` or the explicit external blocker.
For multi-session installation or performance acceptance, verify process count before and after concurrent calls: one shared MCP child with the same PID, not one child per host. Check OS parent processes as well as UXC sessions, because eager native children are invisible to UXC.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |